Multimodal AI for Monitoring and Predicting Neurocognitive Impairment in People with HIV

About This Grant

Abstract/Summary Advances in antiretroviral therapy (ART) have reduced the incidence of severe clinical neurocognitive complications associated with chronic HIV infection, such as HIV-associated dementia (HAD). Nevertheless, nearly half of people with HIV (PWH) still experience asymptomatic neurocognitive disorder (ANI) and mild neurocognitive disorder (MND). Opportunities for using novel, data-driven approaches, such as Artificial Intelligence (AI) in making predictions, real-time monitoring, or improving clinical decision-making to address HIV-related neurocognitive disorders (HAND) proliferate but have yet been fully realized. Recent studies have employed machine learning (ML) and/or deep learning (DL) techniques to either cluster neurocognitive phenotypes or identify key predictors of neurocognitive impairment in PWH. Data from these studies, however, are typically “siloed” and unimodal (e.g., only electronic health records [EHR] data or imaging data). Given the broad spectrum of modalities of neurocognitive disorder, multimodal approach (i.e., integration of different data modalities) provides opportunities to increase robustness and accuracy of diagnostic and prognostic models by utilizing complementary and supplementary information in modalities. However, such multimodal approach is limited often due to the lack of multimodal data and advanced methodologies such as multimodal AI. One novel and ambitious initiative funded by the NIH to advance precision medicine is the All of Us (AoU) Research Program, a centralized data repository, offering secure access to de-identified multimodal data (e.g., EHR data, genomic data, survey data, and imaging data) from almost one million program participants. In our preliminary study, we have developed a computational phenotyping that identified 6,664 confirmed PWH among 633,000+ participants as of October 2023. In response to RFA-MH- 26-105, we propose to apply multimodal AI with a series of longitudinal EHR data (laboratory and medication), genomic data, self-reported survey data (e.g., lifestyle, physical measurement, healthcare access), and imaging data in AoU to 1) identify different biotypes of neurocognitive disorders in PWH (e.g., ANI, MND, HAND) and employ ML/DL approaches to cluster neurocognitive phenotypes; 2) develop, evaluate, and validate multimodal AI models to predict neurocognitive disorders in PWH accounting for comprehensive information and enhance the model interpretability through synergistic integration of a domain-specific knowledge graph; and 3) develop a multimodal AI based decision-making prototype to assist with the identification of PWH with risk of neurocognitive disorders and pilot test its feasibility, usability, and implementation strategies in clinical settings. Personalized risk prediction through multimodal AI could improve the predictive accuracy and early detection of neurocognitive decline in PWH and inform tailored intervention and treatment for PWH. The insights gleaned from our project could also be a demonstration of the power of cutting-edge multimodal AI models to expand our capacity to accelerate HIV care and address the dynamic, complex, and evolving HIV epidemic.